We present **LLaDA-o**, an effective and length-adaptive omni diffusion model for multimodal understanding and generation. LLaDA-o is built on a Mixture of Diffusion (MoD) framework that decouples discrete masked diffusion for text understanding and continuous diffusion for visual generation, while coupling them through a shared, simple, and efficient attention backbone that reduces redundant computation for fixed conditions. Building on MoD, we further introduce a data-centric length adaptation strategy that enables flexible-length decoding in multimodal settings without architectural changes. Extensive experiments show that LLaDA-o achieves strong performance on multimodal understanding and generation benchmarks, reaching 87.04 on DPG-Bench and 0.86 on GenEval. To support future work on omni-diffusion models, we release the model and training/evaluation code at https://github.com/ML-GSAI/LLaDA-o.